A drop-in Vulkan backend for LM Studio that works on old AVX1-only CPUs (Ivy Bridge, Sandy Bridge, older AMD) where the official AVX2 backends refuse to load.
Confirmed loading Gemma 4 and Qwen 3.5 on Ivy Bridge under LM Studio 0.4.14.
- Download the zip from
Releases/GPU Vulkan AVX1/. - Extract it into
C:\Users\%USERNAME%\.lmstudio\extensions\backends\. - Fully quit LM Studio (system tray → Quit) and reopen it.
- In the runtime/engine selector, pick the new backend, then load your model.
If the model fails to load: back up your backends\ folder somewhere safe, then delete everything inside it and extract our zip again so it's the only backend present. LM Studio sometimes picks the wrong runtime when multiple Vulkan backends are installed.
See Releases/GPU Vulkan AVX1/readme.md for full install steps.
See Generate Backends/ for build scripts. The Windows Vulkan script is already configured for AVX1 — just install Visual Studio 2022 Build Tools (with the C++ workload), CMake, Ninja, and the Vulkan SDK, then run the scripts from an x64 Native Tools Command Prompt.
This is a fork of theIvanR/lmstudio-unlocked-backend by theIvanR — all build scripts, the original backend patches, and the foundational work are theirs. This fork adds an updated Vulkan+AVX1 backend built against current llama.cpp main, so newer models like Gemma 4 and Qwen 3.5 load correctly.
Underlying engine: ggml-org/llama.cpp (MIT). The original AVX2-gated binaries belong to LM Studio.
MIT, matching the upstream llama.cpp and the original repo. Unofficial / community-made; use at your own risk and back up your existing LM Studio backends folder before installing.